Answer
Should research and implementation be handled by different agents?
Usually different sessions rather than different agents. The reason is context shape, not specialisation.
Separate the sessions rather than the agents. Investigation reads widely and implementation should read narrowly, so carrying the investigation's context into the change dilutes it. The benefit comes from the context split, not from specialisation.
The case for separation is real and the usual explanation for it is wrong. It is not that a research-shaped agent and an implementation-shaped agent are better at their respective jobs; the same system does both. It is that the two activities want opposite context profiles. Investigation reads broadly and accumulates a lot of material, most of which is navigation. Implementation needs the few conclusions and should not be carrying the search that produced them.
That points at sessions rather than at agents. Run the investigation, have it produce a written finding — what is true, where, with references — and start the implementation from that finding rather than from the session that produced it. Nothing about the second run needs to be specialised; it needs to be uncluttered, and it gets that by starting fresh.
The written finding is the artefact that makes the split work, and its quality determines everything downstream. It should state conclusions with their evidence, name what was ruled out, and flag what remains uncertain. A finding that is a narrative of the investigation defeats the purpose, because reading it costs what carrying the session would have cost.
There is a second reason for the split that is about judgement rather than context. An investigation that flows straight into implementation tends to implement the first workable approach it found, because that approach is present and the alternatives were discarded along the way. A break between the two creates a natural point to ask whether the conclusion is right before building on it, and that question is much cheaper to ask before the code exists.
The case against splitting is that some knowledge does not survive the handover. The investigating run holds impressions that never make it into the finding — that a particular area is messy, that two things are related in a way not worth writing down — and the implementing run works without them. Where the work is small and the investigation short, keeping it in one session is simpler and loses nothing.
So the practical rule is proportional. Short investigation followed by a small change: one session. Substantial investigation, or an investigation that read widely, or a change that will be reviewed by someone else: split, with a written finding in between. The threshold is roughly whether the investigation produced material that the implementation would be better off not carrying.
The reason to split investigation from implementation is that one of them should forget almost everything the other read.
Siddharth Sharma, Context Theory
Related questions
Does the implementing run need to trust the finding?
It should verify the parts the change depends on directly, which is why references matter. A finding that says the value is set in a named file and line can be confirmed in seconds; one that says the configuration is handled centrally cannot, and the implementing run will either re-investigate or proceed on an assumption.
Is this the same as having a planning agent?
Related and not identical. A plan says what to do; a finding says what is true. The distinction matters because a plan produced before the facts are established encodes assumptions, and a finding can be checked independently of any plan. Where both exist, the finding should come first and the plan should reference it.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
2026 speed-to-lead benchmark · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Investigation and implementation require opposite context profiles — broad accumulated reading against a small set of conclusions — so the benefit of separation comes from the context split rather than from any specialisation of the agent. | Comparing the material read during an investigation against the material relevant to the change it produces. |
| Response | A written finding stating conclusions with references, ruled-out options and remaining uncertainty is what makes the split work, whereas a narrative of the investigation costs as much to read as carrying the session would have. | Comparing the length of a finding against the session that produced it and against the conclusions it contains. |
| Software | An investigation flowing directly into implementation tends to build the first workable approach found, because that approach remains present while alternatives were discarded, so the break creates the only cheap point to question the conclusion. | Checking whether the implemented approach was the first viable one identified during the investigation. |
| Constraint | Unwritten impressions held by the investigating run — that an area is messy, that two things relate in a way not worth recording — do not survive the handover, which is why short investigations paired with small changes lose nothing by staying in one session. | Comparing what the investigating run knew against what its written finding contains. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one